Annex III Creates a Presumption

The AI Act has two routes to high-risk classification. The first runs through Article 6(1): your AI is a safety component of a product that EU law already regulates, such as machinery, medical devices or lifts, and that product needs third-party conformity assessment. The second runs through Article 6(2): your system’s intended purpose falls under one of the use cases listed in Annex III, the list that includes recruitment, creditworthiness, education, essential services and biometrics.

A listing in Annex III does not settle the classification, and almost every summary leaves that out. Article 6(3) says an Annex III system “shall not be considered to be high-risk” where it does not pose a significant risk of harm to health, safety or fundamental rights, including by not materially influencing the outcome of decision-making. That derogation determines whether you must meet the full high-risk regime or document an assessment and file a database entry.

If your system is in Annex III, the question is therefore whether you can rebut the presumption and what that costs you. I take the four risk tiers as known; this article deals with the classification decision.

The Article 6(3) Filter: A Chapeau and Four Conditions

The filter is often described as “four exemptions.” That description hides the structure: the four conditions sit under a chapeau and do not stand alone.

The chapeau reads: the system must not pose a significant risk of harm to health, safety or fundamental rights, including by not materially influencing the outcome of decision-making. Four conditions hang off it, and at least one of them must apply. The system is intended to perform a narrow procedural task; or to improve the result of a previously completed human activity; or to detect decision-making patterns or deviations from prior patterns without being meant to replace or influence the previously completed human assessment absent proper human review; or to perform a preparatory task to an assessment relevant to an Annex III use case.

On the wording, one of the four conditions is enough: the first subparagraph applies “where any of the following conditions is fulfilled.” I would not rely on that. A “preparatory task” that materially shapes what the human decides afterwards sits badly with the chapeau, and an authority will ask that question. This is my reading of the text, not a settled interpretation.

Annex III use case? Profiling of natural persons? No significant harm + one of the four? Art. 6(3) Not high-risk but document + register Art. 6(4) · 49(2) High-risk — full Annex III duties risk management, data governance, logging, oversight no yes yes — always no
The classification decision for an Annex III system. Profiling leads straight to “high-risk”, whatever the four conditions say. The “not high-risk” exit still requires a documented assessment and a registration.

Profiling overrides all four conditions

After the list comes a sentence that is easy to miss: notwithstanding the conditions, an Annex III system is always high-risk where it performs profiling of natural persons. Profiling is not weighed against the four conditions; it overrides them.

The AI Act takes the term profiling from the GDPR, whose definition is broad: automated processing of personal data to evaluate personal aspects of a person, particularly to analyse or predict performance at work, economic situation, reliability, behaviour and so on. A tool that ranks CVs can hardly be classed any other way. If your system is in Annex III and evaluates people, the filter is not available to you.

“Narrow” is meant narrowly

The Commission published draft guidelines on high-risk classification on 19 May 2026, and the consultation closed on 23 June 2026. They are not binding, and as far as I can see no final version has been adopted as of October 2026, but market surveillance authorities will lean on them. According to the published analyses of the draft, the Article 6(3) conditions are to be read narrowly.

“Narrow procedural task” means work such as converting a data format, sorting incoming documents into predefined categories or flagging duplicate records: limited work that does not evaluate, score or rank what it processes. If your “narrow procedural task” assigns a fit score, it is neither narrow nor procedural. Anyone who reads the conditions generously here carries a risk they have not priced in.

The Exemption Requires Documentation and Registration

The filter does not take the system out of the AI Act, although many assume it does. Article 6(4) requires this: a provider who concludes that an Annex III system is not high-risk must document that assessment before the system is placed on the market or put into service, must comply with the registration obligation in Article 49(2), and must hand the documentation to national authorities on request.

Whoever claims the exemption therefore opens a file before launch; an inspection is too late to start justifying it. If your reasoning is thin, the authority then has it in writing.

The Commission’s Digital Omnibus on AI proposed removing the registration duty for systems exempted under Article 6(3), and that proposal fell during negotiations. The package agreed on 7 May 2026 and given the Council’s final green light on 29 June 2026 keeps the registration obligation and simplifies the procedure: some fields of Annex VIII, Section B fall away, but the entry stays. Anyone who had planned on the basis of the draft has to register after all.

What the filter replaces. The Article 6(3) filter replaces the full high-risk regime of risk management, data governance, technical documentation, logging, human oversight and conformity assessment with a documented classification assessment and a database entry. That pays off, but it does not free you of everything, and you cannot decide it after launch.

When You Are the Provider

Article 6(4) places the duty on the provider. Most mid-sized companies reasonably consider themselves deployers: they buy the tool, someone else built it, so someone else carries the provider duties. That holds often and fails as often, and projects founder on the cases where it fails.

Article 25 reverses the role. A deployer, distributor or importer becomes the provider of a high-risk system if they put their own name or trademark on it, if they make a substantial modification to it, or if they modify its intended purpose so that it becomes high-risk under Article 6. If you commission a bespoke application-screening tool and run it under your own brand, you provide your own product and no longer deploy someone else’s. The Article 6(3) assessment, the documentation and the registration then fall to you.

The third trigger is the least conspicuous. A general-purpose model or an off-the-shelf tool that was never high-risk becomes high-risk when you use it for an Annex III purpose. Your vendor did nothing wrong and changed nothing; the classification changes because of what you decided to use the tool for. In my experience the works council conversation arrives at the same moment, for the same reason.

The Classification Is an Architecture Decision

How the system is built, not how the project is described, usually decides whether you land inside or outside Annex III’s high-risk tier.

A constructed example shows this. A supplier gets far more applications than its small HR team can read and wants to use AI. Two different systems can come out of that one sentence.

Design A: the system reads each CV, assigns a fit score and ranks the applications by it, so that HR starts at the top. This is Annex III recruitment: the system evaluates people, performs profiling and materially influences the outcome. It is therefore high-risk, and the filter is not available. It remains legal, but you are then building a conformity-assessed system with risk management, data governance and logged human oversight, which takes a different budget and a different timeline.

Design B: the system transfers structured fields from each CV into the applicant tracking system (name, qualifications, languages, years in role), flags duplicate applications and converts attachments to a common format. It does not score and does not sort by suitability. Humans still read every application in the same order as before; only the retyping is gone. That is a narrow procedural task that evaluates nobody and does not shape the decision.

The use case, the department and the budget line are the same, and the classification comes out opposite. The difference lies in the choice to output fields instead of a score, not in a legal opinion.

What I would not do. I would not build a system whose classification depends on a generous reading and hope that the reading holds. Two positions are defensible: plainly outside the high-risk tier by design, or high-risk and built for it. The middle gets expensive, namely a ranker that the documentation calls a “preparatory task” and that holds only until someone reads the code. If a classification holds only on a charitable reading, treat the system as high-risk and budget accordingly.

The Postponement to December 2027

The Digital Omnibus — Regulation (EU) 2026/1744, in force since 27 July 2026 — pushed the high-risk obligations back. For stand-alone Annex III systems they now apply from 2 December 2027; high-risk AI embedded in products regulated under Annex I follows on 2 August 2028. The same package also narrowed “safety component,” so AI that only assists users, optimises performance or handles non-safety quality control does not become high-risk merely by sitting inside a regulated product, unless its failure would endanger health or safety. Unchanged, the prohibitions have applied since February 2025 and the general-purpose AI obligations since August 2025.

If you operate high-risk systems, you should use the extra 16 months. There is also a grace period: under Article 111(2) as amended, the Regulation applies to high-risk systems placed on the market or put into service before the application date only if, from that date, they are subject to significant changes in their design; systems intended to be used by public authorities have their own deadline of 2 August 2030. A system shipping in 2026 that scores applicants therefore does not fall under the high-risk obligations on 2 December 2027 by itself.

The classification question still cannot wait, because the design decisions you make now fix the classification, and the grace period lasts only as long as you do not significantly change the system’s design. A production system you keep developing loses it with the first significant change. You then retrofit risk management, data governance and logged oversight into a live system, which is the most expensive route to compliance. If you have ever had to reconstruct what a model did many months ago, you also know that the record-keeping has to be designed in from the start. So settle the classification in this quarter’s architecture review, not shortly before December 2027.